Research On Bladder CT Contour Anomaly Detection Based On Virtual Sample Generation And Feature Fusion

The contour of the organs drawn by the physician based on the patient’s CT image determines the location and dose of radiation therapy and plays an essential role in radiation therapy planning.However,differences in mapping levels are one of the main sources of uncertainty in radiotherapy,and poor organ contours can lead to wrong radiation doses and medical malpractice.Therefore,detecting abnormal organ contours is necessary to ensure the curative effect.This part of the work is mainly completed through peer review.Although it can effectively reduce the risk of medical malpractice,it increases the burden on the medical system.In some medical resources,in places where there is a shortage,it is not feasible,so to decompress the medical system and reduce the workload of doctors,this paper takes the bladder contour as the research object and researches the automatic detection method of abnormal contour.The work results are as follows:(1)A virtual sample generation method based on geometric features was proposed to solve the unbalanced number of data sets.First,the shape and volume of real rare samples of two types of Discontinuities and Surface aberrations were statistically analyzed.Then,we constructed the virtual sample generation algorithms of these two kinds of samples according to the extracted knowledge to generate a sufficient number of samples and balance the number of various samples in the dataset.(2)A kind of multi-view convolutional neural networks(MV-CNNs)method was proposed based on the double-layer fusion of view and modal layers to recognize CT bladder contour.To provide more information for model training,we used images of three modes(original image,LBP image,and Canny operator image)as input.A weight adaptive joint pooling method was proposed to make full use of the information of each view feature to generate modal features at the view layer.In addition,the Concatenation/Addition method was used to merge three modal features to generate shape features at the modal layer for model recognition.Compared with other traditional deep learning models such as MV-CNNs,3D-CNN,DGCNN,and RANDAL-NET,our proposed MV-CNNs method based on the double-layer fusion of view layer modal layer can reach the highest testing accuracy of 93.1%,showing the effectiveness of the proposed method.